Anomaly Detection for Data Streams Based on Isolation Forest Using Scikit-Multiflow

Togbe, Maurras Ulbricht; Barry, Mariam; Boly, Aliou; Chabchoub, Yousra; Chiky, Raja; Montiel, Jacob; Tran, Vinh-Thuy

Detecting anomalies in streaming data is an important issue in a variety of real-word applications as it provides some critical information, e.g., Cyber security attacks, Fraud detection or others real-time applications. Different approaches have been designed in order to detect anomalies: statistics-based, isolation-based, clustering-based. In this paper, we present a quick survey of the existing anomaly detection methods for data streams. We focus on Isolation Forest (iForest), a state-of-the-art method for anomaly detection. We provide the implementation of IForestASD, a variant of iForest for data streams. This implementation is built on top of scikit-multiflow, an open source machine learning framework for data streams. In fact, few anomalies detection methods are provided in the well-known data streams mining frameworks such as MOA or StreamDM. Hence, we extend scikit-multiflow providing an additional tool. We performed experiments on 3 real-world data sets to evaluate predictive performance and resource consumption (memory and time) of IForestASD and compare it with a well known and state-of-the-art anomaly detection algorithm for data streams called Half-Space Trees.